Publications

Retry mechanisms can prolong a service overload even after a temporary capacity disruption has been resolved. Although restricting retries may help service recovery, that recovery does not necessarily reflect the proportion of users who receive successful responses. This paper introduces RetryCascade, a reproducible discrete-event simulator that distinguishes logical requests from physical retry attempts and evaluates outcomes by request-arrival period. An exploratory evaluation consists of 630 simulation runs covering varying workload intensities, independent pre-admission failures, transient capacity reductions, and client response policies. The primary comparison examines the effect of introducing a shared retry budget while keeping exponential backoff and full jitter, compared with a no-retry baseline. An additional validation experiment comprises 60 runs using ten previously unseen random seeds under a fixed experimental condition. Both the retry-budget and no-retry configurations achieved recovery in all ten runs that had active triggers. However, their average success rates for requests arriving during the incident differed substantially: 38.27% for the retry-budget configuration and 65.55% for the no-retry configuration. For requests arriving within the subsequent 60-second period, the corresponding success rates were 98.51% for the retry-budget configuration and 89.87% for the no-retry configuration. A separate exploratory sensitivity analysis further shows that waiting for outstanding requests to complete and accepting delayed responses can significantly influence these results. Overall, the findings highlight the importance of evaluating service recovery together with arrival-cohort success rates and clearly defined client response semantics. These conclusions are based on a synthetic simulation environment and should not be interpreted as direct predictions of production database behavior.

The evolution of large language models (LLMs) into agentic artificial intelligence systems capable of autonomous decision-making has created a structural transformation in the AI ecosystem. While agents' dynamic interaction with the external world provides exceptional flexibility, it also introduces unique agent-specific threat vectors that traditional security approaches do not cover. This article examines the security risks faced by LLM-based autonomous agents and their defense strategies through a holistic framework spanning the data, model, and usage layers. Critical vulnerabilities such as indirect prompt injection, tool hijacking, and contagious attacks in multi-agent environments are analyzed. Mitigation strategies including Secure by Design architectures, authorization mechanisms, structured query validation, and Human-In-The-Loop (HITL) controls are presented. The study also examines the concept of Cyber Threat Inflation caused by autonomous systems and emphasizes that agents should be positioned as proactive cyber defenders in future cybersecurity architectures.

In recent years, artificial intelligence has evolved from being merely an outcome of technological innovation into a strategic field with a direct impact on economic competitiveness, the effectiveness of public administration, and national security policies. In this context, the widespread adoption of AI technologies cannot be regarded as a process limited solely to increasing technical capacity. Unless these technologies are addressed within a holistic framework encompassing security, ethics, data governance, and skilled human-resource policies, structural and managerial risks may emerge in the long term. This strategic perspective demonstrates that artificial intelligence is critically important not only economically but also in terms of public safety and social order.

With the advancement of artificial intelligence (AI) technologies, generative AI models are delivering groundbreaking innovations in many areas, including natural language processing (NLP), image generation, code generation, and speech synthesis. Large language models (LLMs), in particular, are widely used for text generation, information retrieval, and decision-support systems. This study examines the fundamental concepts of generative AI models, comparison criteria, pricing structures, and API/chat usage models.

Recent advances in artificial intelligence (AI) have highlighted the importance of systems that operate more independently and flexibly than traditional narrow-task systems. In this context, Agentic AI refers to autonomous AI systems capable of defining their own goals without human intervention and adapting to environmental variables to achieve those goals. Through flexible decision-making mechanisms and the ability to continue operating in dynamic and complex environments, Agentic AI systems offer significant potential in areas such as healthcare, finance, and disaster management. This study demonstrates that Agentic AI is not merely a technical innovation but a multidimensional transformation that must be integrated with society, ethics, and governance. For these systems to succeed, engineering capabilities must be developed in parallel with ethical safeguards, user expectations, transparency standards, and legal frameworks.

The healthcare sector directly affects individuals' quality of life while also requiring new solutions because of the continuously increasing volume and complexity of data. Artificial intelligence (AI) technologies therefore represent not only an innovation but also a key factor in redefining healthcare services. Through their capacity to process massive healthcare datasets, recognize patterns, and learn from data, AI systems provide solutions in areas such as early disease diagnosis, personalization of treatment plans, enhancement of clinical decision-support systems, and automation of administrative processes. Deep-learning architectures, in particular, can manage this complexity and generate highly accurate predictions. This study comprehensively examines the use of artificial intelligence models in healthcare and their future prospects.

The 21st century is witnessing one of the most important turning points in the history of artificial intelligence. Large language models (LLMs) are among the principal components of this technological transformation. Beyond using natural language processing (NLP) techniques solely for information extraction, these models can perform numerous complex tasks such as text generation, reasoning, translation, coding, multilingual communication, and interactive reasoning. Models including the Generative Pre-trained Transformer (GPT) series (OpenAI), Bidirectional Encoder Representations from Transformers (BERT) and PaLM (Google), Claude (Anthropic), Gemini (DeepMind), LLaMA (Meta), Mistral, Gemma, Qwen, and DeepSeek have gained historical significance not only through their technical performance but also through their influence on academic, social, and political agendas. LLMs are increasingly relevant not only to software engineers and data scientists but also to lawyers, teachers, physicians, and public administrators. Against this backdrop, this study comprehensively compares the technical characteristics and application potential of Meta's Large Language Model Meta AI (LLaMA) series (3.1, 3.2, 3.3, and 4) with open-source language models.

As artificial intelligence (AI) systems become more complex, the need to make their decision-making processes transparent and understandable has increased. The use of powerful but "black-box" models such as deep learning can result in decisions that are difficult to interpret, particularly for complex structures such as time-series data. Interest in explainable artificial intelligence (XAI) methods is therefore increasing to clarify why decisions are made and to build user trust. Explainability makes the logical basis of an AI model's decisions understandable, while transparency refers to the ability to provide information about the model's internal operation. Accountability requires mechanisms through which responsibility for the consequences of these decisions can be assigned. Together, these three elements support not only technical success but also ethical responsibility. XAI methods reveal model decision logic through techniques such as visualization, feature importance, and impact analysis. Such explanations are important for building user trust, identifying model errors, and validating decision processes. Some methods have also been adapted for non-traditional data types such as time series. In this context, the use and effectiveness of explanatory techniques such as LIME, SHAP, and Saliency Maps across different data types have been investigated in various studies.

In recent years, artificial intelligence has had a significant impact across a broad spectrum ranging from everyday life to industrial production, education and healthcare, public administration, and defense technologies. Systems capable of interacting with text, images, audio, and unstructured data are making functions such as search, summarization, recommendation generation, decision support, and automation increasingly reliable and accessible. This transformation is not limited to the emergence of new tools; it is also fundamentally changing how institutions access, verify, and integrate information into business processes. Particularly in technical domains, generating answers grounded in verified sources has become not merely useful but essential. One approach addressing this requirement is Retrieval-Augmented Generation (RAG), which connects generative systems to institutional or domain-specific knowledge sources so that generated text is grounded in documents. Rather than relying only on the model's high-level knowledge, RAG retrieves document passages semantically related to the user's query and constructs the answer based on those passages. This creates a practical bridge between fluent text generation, traceability, and source grounding, reducing hallucination risk and improving domain-specific accuracy. This study focuses on the comparative analysis of different language models supported by RAG under the same domain documents and experimental conditions. Responses generated by different language models are examined within a single, consistent RAG pipeline built on STM32F103C8 technical documentation. The evaluation emphasizes source-grounded accuracy, technical specificity, and consistency rather than merely fluent text. By using the same document pool, contextualization approach, and standardized question set, the comparison aims to provide a fair and reproducible framework for assessing model capabilities. The methodology section presents the dataset and experimental setup in detail; the findings summarize question-level and model-level observations; and the discussion addresses practical implications and limitations. Overall, the study presents a domain-specific, application-oriented evaluation approach for the common engineering requirement to rapidly locate correct information and generate clear, source-grounded answers.